Part 2: Mangrove Extent Mapping and Time Series•Google Earth Engine Asset •External dataset...

National Aeronautics and Space Administration

Abigail Barenblitt & Temilola Fatoyinbo

Nov 12th, 2020

Part 2: Mangrove Extent Mapping and Time Series

2NASA’s Applied Remote Sensing Training Program

Course Structure and Materials

• Three, 1.5-hour sessions on

November 5, 12, and 19

• The same content will be presented

at two different times each day:

– Session A: 10:00-11:30 EST (UTC-5)

– Session B: 15:00-16:30 EST (UTC-5)


Course Structure and Materials

• Webinar recordings, PowerPoint presentations, and the homework

assignment can be found after each

session at:

– https://appliedsciences.nasa.gov/join-mission/training/english/remote-sensing-

mangroves-support-un-sustainable-

development-goals

– Q&A following each lecture and/or by

email at:

• [email protected] or

• [email protected]

https://appliedsciences.nasa.gov/join-mission/training/english/remote-sensing-mangroves-support-un-sustainable-development-goals

mailto:[email protected]



Homework and Certificates

• Homework:

– Three homework assignments, assigned after each weekly Part

– Answers must be submitted via Google Forms

• Certificate of Completion:

– Attend all three live webinars

– Complete the homework assignments by the deadline (access from

ARSET website)

– You will receive certificates approximately two months after the

completion of the course from: [email protected]



Prerequisites

• Required Version of QGIS: 3.10 https://www.qgis.org/en/site/forusers/download.html

• Download and Install Class Accuracy Plug-in for QGIS: https://github.com/remotesensinginfo/classaccuracy

• For instructions for installation refer to this video: https://www.youtube.com/watch?v=NJRdKpmujRo

• Fundamentals of Remote Sensing

• Intro to JavaScript for GEE

• Create a Google Earth Engine Account

Optional:

• GEE Beginner’s Cookbook

• GEE Managing Assets

• Introduction to Google Earth Engine Tutorial

https://www.qgis.org/en/site/forusers/download.html

https://github.com/remotesensinginfo/classaccuracy

https://www.youtube.com/watch?v=NJRdKpmujRo

https://appliedsciences.nasa.gov/join-mission/training/english/fundamentals-remote-sensing

https://developers.google.com/earth-engine/tutorial_js_01

https://developers.google.com/earth-engine/tutorials/community/beginners-cookbook

https://developers.google.com/earth-engine/asset_manager

https://www.youtube.com/watch?v=uHtehTSw7vg


Learning Objectives

By the end of this presentation, you will:

• Understand the basics of using Google Earth Engine

• Create a mangrove extent map using a Random Forest Classification

• Create a time series for mangrove extent change


Outline

1) Review of Google Earth Engine

2) Review of Time Series Analysis

3) Demo of Time Series Analysis for Guyana

– Setting Up the Map

– Filtering a Landsat Composite

– Constructing a Random Forest Model

– Time Series Comparison

– New Random Forest

– Calculating Mangrove Area

– Exporting Layers of Interest

– Demo of QGIS Class Accuracy Plug-in


What is Google Earth Engine?

• A free, open-source, “cloud-based geospatial processing platform”

• Comprised of:

• A catalog of publicly available datasets

• Google’s computation power

• An Application Programming Interface (API)

• A code editor

Google’s Mission:


How GEE Works

• Cloud-Based

– Client vs. server functions

– Users manipulate “proxy” objects through the server

– Instructions sent to Google for processing and results sent back to web

browser for display

• Defaults to WGS84 Projections

• Capabilities/Limitations

– Planetary-scale

– Quota restrictions due to open-source nature

– User memory limited


The Code Editor

• For more detailed analyses

• JavaScript code editor (Python available)

• Map display

• API reference documentation (Docs tab)

• Console output (Console tab)

• Task Manager (Tasks tab)

• Interactive map query (Inspector tab)


Glossary of Terms

• Google Earth Engine Asset• External dataset loaded into Google Earth Engine for analysis

• Table• Vector data in shapefile format• Example: Ground-truthed location data

• Image• Raster data composed of one or more bands• Example: Euclidean distance to stream

• Image Collection• A stack or time series of images• Example: Landsat 8 imagery


Resources for Help

• Docs Tab

• Developer’s Guide

• Google Earth Engine Developers Group

https://developers.google.com/earth-engine

https://groups.google.com/forum/#!forum/google-earth-engine-developers

Time Series of Mangrove Extent


Understanding Change Over Time

• We can use a time series analysis to understand how mangroves have changed

over time

• Identify areas of loss/gain

• Understand patterns of change


Time Series Analyses

• Examine the same study area over several years

• Focus on specific years of interest (ex: decadal info)

• Example: Comparison of mangrove extent in Colombia, 2000 vs. 2019

2000 2019


Using Landsat to Compare Extent

2000 2019



• We can compare the values of

different indices like Normalized

Difference Vegetation Index (NDVI)

across different years.

• Higher values of of NDVI indicate

higher levels of vegetation

(mangroves).

2000



• We can create samples of areas with

and without mangroves using this

imagery.

• Machine learning allows us to use

these samples to detect mangroves

across a region.


Random Forest Classification

• Machine Learning

–Uses statistics to identify patterns in

large datasets

–AI that “learns” from data

• Ensemble, tree-based Machine Learning Algorithm

• Supervised

• Uses decision trees to select best solution by “voting”


The Black Box

Input and

Training Data

Output

Machine

Learning

Algorithm

Machine

Learning

Algorithm

The machine learning

algorithm is a “black

box,” where we can

only observe the

inputs and outputs.


Random Forest Classification

• The algorithm constructs a decision tree for each sample.

• Based on the predictors (bands from Landsat), the trees will vote for each pixel to detect mangrove vs. non-mangrove.

• The most supported value is assigned to each pixel.


Mangrove Extent Using Random Forest Classification

1) Create Samples 2) Run RFC 3) Refine Model


Validation of Results

• After we have created our map, we need to validate it.

• Understanding the accuracy of our model allows us to understand how reliable

our results are.

• We can use stratified random samples to run an independent accuracy

assessment.

• We ”visit” each point and use satellite imagery to mark if they are correct.

• For this exercise, we will use the Class Accuracy Plug-in in QGIS 3.10:

– https://github.com/remotesensinginfo/classaccuracy



Run Model

Assess Accuracy

Refine Model

Input Training

Data

Final Model


Study Area: Guyana

• NASA and SERVIR work with in-

country partners in countries like

Guyana to help them monitor

mangroves.

• Guyana foresees future flooding

and saltwater intrusion as sea levels

rise.

• This case study shows how mapping

mangroves can help focus

conservation practices.


TO FOLLOW ALONG WITH THE FULL SCRIPT:

https://code.earthengine.google.com/a2400e2ce048914ccf1b16aba2702951

https://code.earthengine.google.com/a2400e2ce048914ccf1b16aba2702951

1) Setting Up the Map


Import the Following

• Landsat 8 Surface Reflectance Tier 1

• Image Collection: LANDSAT/LC08/C01/T1_SR

• Rename “L8”

• SRTM Digital Elevation Data 30m

• Image Collection: USGS/SRTMGL1_003

• Rename “SRTM”


Start by Setting Up the Map

Center the map to the region of interest using the region shapefile.

Map.centerObject(geometry,7)

Map.setOptions('satellite')


Draw a Geometry for Area of Interest

2) Setting Up a Filtered Landsat Composite


Set Up a Filtered Landsat Composite

First, we need to mask clouds.

Landsat data includes a 'pixel_qa' band which can be used to create a function to mask clouds.

function maskClouds(image) {

// Bits 3 and 5 are cloud shadow and cloud, respectively.

var cloudShadowBitMask = ee.Number(2).pow(3).int();

var cloudsBitMask = ee.Number(2).pow(5).int();

// Get the pixel QA band.

var qa = image.select('pixel_qa');

// Both flags should be set to zero, indicating clear conditions.

var mask = qa.bitwiseAnd(cloudShadowBitMask).eq(0).and(qa.bitwiseAnd(cloudsBitMask).eq(0));

// Return the masked image, scaled to [0, 1].

return image.updateMask(mask).divide(10000).copyProperties(image, ["system:time_start"]);

}


Spectral Indices

• NDVI = Normalized Difference Vegetation Index (Red and NIR)– Quantifies Vegetation

• NDMI = Normalized Difference Moisture Index (NIR and SWIR)– Vegetation Water Content

• MNDWI = Modified Normalized Difference Water Index (Green and SWIR)– Water Information

• SR = Simple Ratio (Red and NIR)– Simple Vegetation Index

• Ratio54 = Band Ratio 54 (SWIR and NIR)– Maps Water Features

• Ratio35 = Band Ratio 35 (Red and SWIR)– Maps Water Features

• GCVI = Green Chlorophyll Vegetation Index (NIR and Green)– Green Leaf Biomass


Add Spectral Indices

This function maps spectral indices for Mangrove Mapping using Landsat 8.

var addIndicesL8 = function(img) {

var ndvi = img.normalizedDifference(['B5','B4']).rename('NDVI’); // NDVI

// NDMI (Normalized Difference Mangrove Index - Shi et al 2016 - New spectral metrics for mangrove forest identification)

var ndmi = img.normalizedDifference(['B7','B3']).rename('NDMI’);

var mndwi = img.normalizedDifference(['B3','B6']).rename('MNDWI’); // MNDWI (Modified Normalized Difference Water ndex -

Hanqiu Xu, 2006)

var sr = img.select('B5').divide(img.select('B4')).rename('SR’); // SR (Simple Ratio)

var ratio54 = img.select('B6').divide(img.select('B5')).rename('R54’); // Band Ratio 54

var ratio35 = img.select('B4').divide(img.select('B6')).rename('R35’); // Band Ratio 35

var gcvi = img.expression('(NIR/GREEN)-1’,{ 'NIR':img.select('B5’), 'GREEN':img.select('B3’)}).rename('GCVI’); // GCVI

return img

.addBands(ndvi)

.addBands(ndmi)

.addBands(mndwi)

.addBands(sr)

.addBands(ratio54)

.addBands(ratio35)

.addBands(gcvi);

};


Filter Landsat Data by Date

// Select the desired central year here

var year = 2019;

// Start date will be set one year before the central year

var startDate = (year-1)+'-01-01';

// End date will be set to one year later than the central year

var endDate = (year+1)+'-12-31';


Apply Filters and Masks to Landsat 8 Imagery

var l8 = L8.filterDate(startDate,endDate)

// Mask for clouds and cloud shadows

.map(maskClouds)

// Add the indices

.map(addIndicesL8)


Composite the Landsat Image Collection

You can composite on a per-pixel, per-band basis using “.median()”

var composite = l8

// Uses the median reducer

.median()

// Clips the composite to our area of interest

.clip(geometry);


Mask to Areas of Low Elevation and High NDVI and MNDWI

Additional masking allows us to focus on areas that are more likely to have mangroves.

// Clip SRTM data to region

var srtmClip = SRTM.clip(geometry);

// Mask to elevations less than 65 meters

var elevationMask = srtmClip.lt(65);

// Used the NDVI and MNDWI bands to create masks

var NDVIMask = composite.select('NDVI').gt(0.25);

var MNDWIMask = composite.select('MNDWI').gt(-0.50);

// Apply the masks

var compositeNew = composite

.updateMask(NDVIMask)

.updateMask(MNDWIMask)

.updateMask(elevationMask)


Display Results

We need to map the Landsat composite to assemble our training data.

// Select bands and parameters for visualization

var visPar = {bands:['B5','B6','B4'], min: 0, max: 0.35};

// Add layer to map

Map.addLayer(compositeNew

.clip(geometry), visPar, 'Landsat

Composite 2019')

3) Constructing the Model


Construct a Random Forest Classification

Use the displayed Landsat data to add polygons of training data.


Prepare Training Data and Predictors

// After drawing training polygons, merge them together

var classes = Mangrove.merge(NonMangrove)

// Define the bands you want to include in the model

var bands = ['B5','B6','B4','NDVI','MNDWI','SR','GCVI']

// Create a variable called image to select the bands of interest and clip to geometry

var image = compositeNew.select(bands).clip(geometry)

// Assemble samples for the model

var samples = image.sampleRegions({

collection: classes, // Set of geometries selected for training

properties: ['landcover'], // Label from each geometry

scale: 30 // Make each sample the same size as Landsat pixel

}).randomColumn('random'); // creates a column with random numbers


Split Samples for Testing

Here we randomly split our samples to set some aside for testing our model's accuracy using the "random" column we created.

var split = 0.8; // Roughly 80% for training, 20% for testing

var training = samples.filter(ee.Filter.lt('random', split)); // Subset training data

var testing = samples.filter(ee.Filter.gte('random', split)); // Subset testing data



Print these variables to see how much training and testing data you are using.

print('Samples n =', samples.aggregate_count('.all'));

print('Training n =', training.aggregate_count('.all'));

print('Testing n =', testing.aggregate_count('.all'));


Begin Random Forest Classification

“.smileRandomForest” is used to run the model. Here we run the model using 100 trees and 5 randomly selected predictors per split ("(100,5)").

var classifier = ee.Classifier.smileRandomForest(100,5).train({

features: training.select(['B5','B6','B4','NDVI','MNDWI','SR','GCVI’, 'landcover']), // Train using bands and landcover property

classProperty: 'landcover', // Pull the landcover property from // classes

inputProperties: bands

});


Test the Fit of the Model

var validation = testing.classify(classifier);

var testAccuracy = validation.errorMatrix('landcover’, 'classification');

print('Validation error matrix RF: ', testAccuracy);

print('Validation overall accuracy RF: ‘,

testAccuracy.accuracy());


Classify the Landsat Composite Using the Random Forest Model

var classifiedrf = image.select(bands) // select the predictors

.classify(classifier);

// .classify applies the Random Forest


Reduce Noise in Results

The model results may be "noisy." To reduce noise, create a mask to maskunconnected pixels.

// Create an image that shows the number of pixels each

// pixel is connected to

var pixelcount = classifiedrf.connectedPixelCount(100, false);

// Filter out all pixels connected to 4 or less

var countmask = pixelcount.select(0).gt(25);


Mask Results

Mask the results to only display mangrove extent.

var classMask = classifiedrf.select('classification').gt(0)

var classed = classifiedrf.updateMask(countmask)

.updateMask(classMask)


Map Results

Mask the results to only display mangrove extent

// Add classification to map

Map.addLayer (classed, {min: 1, max: 1, palette:'blue'}, 'Mangrove Extent 2019');

// For comparison, let's add the GMW dataset to the map.

var GMW = ee.Image('projects/mangrovescience/GuyanaGMW')

Map.addLayer (GMW, {min: 1, max: 1, palette:'green'}, 'Global Mangrove Watch');

4) Time Series Comparison


Time Series Comparison

We want to be able to compare mangrove extent in different years to

examine if mangrove area has been lost over time.

We need to rerun our model using Landsat imagery from a different date of

interest with new training data.


Adding Landsat 7 Spectral Indices

• Landsat 5 and 7 have different band numbers.

• Landsat 8 also collects images in wider swaths.

• We will need to assign spectral indices using different band values.


Spectral Indices for Landsat 8


Spectral Indices for Landsat 7


Adding Landsat 7 Spectral Indices

var addIndicesL7 = function(img) {

var ndvi = img.normalizedDifference(['B4','B3']).rename('NDVI');

var ndmi = img.normalizedDifference(['B7','B2']).rename('NDMI');

var mndwi = img.normalizedDifference(['B2','B5']).rename('MNDWI');

var sr = img.select('B4').divide(img.select('B3')).rename('SR');

var ratio54 = img.select('B5').divide(img.select('B4')).rename('R54');

var ratio35 = img.select('B3').divide(img.select('B5')).rename('R35');

var gcvi = img.expression('(NIR/GREEN)-1',{ 'NIR':img.select('B4’), 'GREEN':img.select('B2')}).rename('GCVI');

return img.addBands(ndvi)

.addBands(ndmi)

.addBands(mndwi)

.addBands(sr)

.addBands(ratio54)

.addBands(ratio35)

.addBands(gcvi)};


Filter Landsat Data by Date and Region

// Select the desired central year here

var year = 2009;

// Start date will be set one year before the central year

var startDate = (year-1)+'-01-01';

// End date will be set to one year later than the central year

var endDate = (year+1)+'-12-31';


Apply Filters and Masks

var l7 = L7.filterDate(startDate,endDate)

// Mask for clouds and cloud shadows

// We use the same function we used for Landsat 8 to mask clouds

.map(maskClouds)

// Add the indices

.map(addIndicesL7)


Composite Image Collection

var L7composite = l7

// Uses the median reducer

.median()

// Clips the composite to our area of interest

.clip(geometry);


Mask to Low Elevation/High NDVI and MNDWI

var L7NDVIMask = L7composite.select('NDVI').gt(0.25);

var L7MNDWIMask = L7composite.select('MNDWI').gt(-0.50);

// Apply the masks

var L7compositeNew = L7composite

.updateMask(L7NDVIMask)

.updateMask(L7MNDWIMask)

.updateMask(elevationMask) //We use the same mask


Display Results

// Select bands and parameters for visualization

// We use bands 4, 5, and 3 instead

var L7visPar = {bands:['B4','B5','B3'], min: 0, max: 0.35};

// Add layer to map

Map.addLayer(L7compositeNew.clip(geometry), L7visPar, 'Landsat

Composite 2009')

5) New Random Forest Model


Construct New Random Forest Model

Prepare training data and predictors.

// After drawing training polygons, merge them together.

var classes2009 = Mangrove2009.merge(NonMangrove2009)

// Define the bands you want to include in the model.

var L7bands = ['B4','B5','B3','NDVI','MNDWI','SR','GCVI']

// Create a variable called “image” to select the bands of interest and clip to

geometry.

var L7image = L7compositeNew.select(L7bands).clip(geometry)


Assemble Samples

var L7samples = L7image.sampleRegions({

collection: classes2009, // Set of geometries selected for training

properties: ['landcover'], // Label from each geometry

scale: 30 // Make each sample the same size as Landsat pixel

}).randomColumn('random'); // creates a column with random numbers



// Subset training data.

var L7training = L7samples.filter(ee.Filter.lt('random', split));

// Subset testing data.

var L7testing = L7samples.filter(ee.Filter.gte('random', split));

// Print these variables to see how much training and testing data you are

// using.

print('Samples n =', L7samples.aggregate_count('.all'));

print('Training n =', L7training.aggregate_count('.all'));

print('Testing n =', L7testing.aggregate_count('.all'));


Begin Random Forest Classification

var L7classifier = ee.Classifier.smileRandomForest(100,5).train({

features: L7training.select(['B4','B5','B3','NDVI','MNDWI','SR','GCVI',

'landcover']), // Train using bands and landcover property

classProperty: 'landcover', // Pull the landcover property from classes

inputProperties: L7bands

});


Classify Landsat Composite

var L7classifiedrf = L7image.select(L7bands) // Select the predictors

.classify(L7classifier);

// Reduce Noise

var pixelcount = L7classifiedrf.connectedPixelCount(100, false);

var countmask = pixelcount.select(0).gt(25);


Map Results

// Mask results to only display mangrove extent

var L7classMask = L7classifiedrf.select('classification').gt(0)

var L7classed=

L7classifiedrf.updateMask(countmask).updateMask(L7classMask)

Map.addLayer (L7classed, {min: 1, max: 1, palette:'green'}, 'Mangrove Extent

2009');

6) Calculate Mangrove Area


Calculate Mangrove Extent

// 2009

var get2009 = L7classed.multiply(ee.Image.pixelArea()).divide(10000).reduceRegion({

reducer:ee.Reducer.sum(),

geometry:geometry,

scale: 100,

maxPixels:1e13,

tileScale: 16

}).get('classification');

print(get2009, 'Mangrove Extent 2009 in ha')


Calculate Mangrove Extent

// 2019

var get2019 = classed.multiply(ee.Image.pixelArea()).divide(10000).reduceRegion({

reducer:ee.Reducer.sum(),

geometry:geometry,

scale: 100,

maxPixels:1e13,

tileScale: 16

}).get('classification');

print(get2019, 'Mangrove Extent 2019 in ha')

7) Create Points for Accuracy Assessment


Running Independent Accuracy Assessments

To test the actual accuracy of the model rather than the fit, we will need to

create random sampling points.

var stratSamples = classed.stratifiedSample({

numPoints:150, // Number of points per class

classBand: 'classification',

region:geometry,

scale: 30,

geometries:true });


Buffer Points

// Add a 15m-radius buffer around each point.

var stratBuff = function(feature) {

var num = feature.get('classification');

return feature.buffer(15).set('classification', num);

};

var stratPoints = stratSamples.map(stratBuff)

8) Export Layers of Interest


Export Layers of Interest

// 2019 Mangrove Extent

Export.image.toDrive({

image: classed,

description: '2019GuyanaMangroveExtent',

region: geometry,

scale: 30,

maxPixels: 1e13

});



// 2009 Mangrove Extent

Export.image.toDrive({

image: L7classed,

description: '2009GuyanaMangroveExtent',

region: geometry,

scale: 30,

maxPixels: 1e13

});



// Stratified Random Samples

Export.table.toDrive({

collection: stratPoints,

description:'StratifiedrandomPoints',

fileFormat: 'SHP',

});

9) QGIS Class Accuracy


Class Accuracy Plug-in

• Class Accuracy is a QGIS Plug-in created by Dr. Peter Bunting.

• This tool takes the user through each randomly stratified point.

• The user then defines whether the point was accurately classed by the

model or not.

• The result is an assessment of the overall accuracy of the model.


QGIS Class Accuracy Plug-In

• Open QGIS 3.10 and add a satellite basemap (ex. Bing Aerial).

• Add exported random points:

– Add two columns: Export and Processed

– Ensure all columns including classification are in String format.

• Open Class Accuracy Plug-in:

– https://github.com/remotesensinginfo/classaccuracy



Recap

During this lesson we covered:

• Mangrove extent mapping over two time periods

• Independent accuracy assessment of the model

Next time we will cover:

• Creating country-specific apps

• Example applications of results


Questions

• Please enter your questions into the Q&A box.

• We will post the questions and answers to the

training website following the conclusion of

the course.


Contacts

• Contacts

– Abigail Barenblitt: [email protected]

– Temilola Fatoyinbo: [email protected]

• General ARSET Inquiries

– Ana Prados: [email protected]

• ARSET Website:

– appliedsciences.nasa.gov/arset

http://[email protected]

http://[email protected]


http://appliedsciences.nasa.gov/arset


Thank You!

Part 2: Mangrove Extent Mapping and Time Series•Google Earth Engine Asset •External dataset...

Documents

Transcript of Part 2: Mangrove Extent Mapping and Time Series•Google Earth Engine Asset •External dataset...